Prediction Horizons in Agent Models

One motivation for many agent-based models is to predict the future. The nonlinearity of agent interactions in most non-trivial domains mean that the usefulness of such predictions will be limited beyond a certain point (the “prediction horizon”), due to unbounded divergence of their trajectories. The model’s predictions are increasingly useful out to the prediction horizon, but become misleading beyond that point. We exhibit and characterize this behavior in a simple model, based on the polyagent modeling construct, which uses multiple ghost agents to explore alternative futures concurrently for a domain entity. We also discuss how a single agent in such a model can estimate the prediction horizon based on locally available information, and use this estimate to modulate dynamically how far it seeks to look into the future.

• Type: Conference
• Year: 2007
• Where: EEMMAS
• Authors: H. Van Dyke Parunak, Theodore C. Belding, Sven A. Brueckner
• Keywords: Artificial Intelligence, EEMMAS 2007, Model’s Predictions, Prediction, Prediction Horizon